| name | assumption-graph-loop |
| description | Autonomous belief-revision loop — fuse @ai: annotations + /deep-research claims into the temporal assumption graph and revise the persistent belief log over time |
Assumption-Graph Loop
One turn of the longitudinal belief loop for the temporal assumption graph
(crates/ix-assumption-graph, contract docs/contracts/2026-05-31-assumption-graph.contract.md):
@ai: annotations ─┐
├─► fuse (ix-fuzzy, Belnap C synthesis) ─► revise ─► belief-events.jsonl
research claims ──┘ cross-source only (append on verdict change)
It ingests the workspace's @ai: annotations plus any research-domain claims,
fuses each claim's evidence, and appends a BeliefEvent for every claim whose
verdict changed since the last run. Re-running on unchanged evidence is a no-op
— so it is safe to schedule.
When to Use
- On a schedule (daily / weekly), to keep the belief log current as code and
research evolve — the "semantic story over a long period".
- After a
/deep-research run, to fold its verified findings into the graph and
detect where new evidence contradicts a standing assumption.
- Before a review, to surface
escalated (Contradictory) claims.
Run It
cargo run -p ix-assumption-graph --bin ix-assumption-graph-loop -- \
--workspace . \
--research state/assumptions/research-claims.json \
--log state/assumptions/belief-events.jsonl \
--trigger deep-research-reverify
Defaults: --workspace ., --log state/assumptions/belief-events.jsonl,
--trigger loop, no research file. The log is created (with parent dirs) if
absent. Output reports node count, total/appended belief events, and each flip
(T -> C, etc.).
The research-claims.json shape
A JSON array. The truth-value judgment — mapping a verified finding to a
hexavalent value — is the adapter's responsibility, NOT the crate's, so this
file is the seam between /deep-research's output and the graph:
[
{
"claim": "p99 latency under 5ms",
"truth_value": "F",
"confidence": 0.85,
"source": "deep-research",
"evidence": "arxiv:2510.11822"
}
]
truth_value: one of T P U D F C (hexavalent).confidence: 0.0–1.0.source (optional, default "deep-research"): the independence class.
Distinct sources are what let fusion synthesize a contradiction — give
genuinely independent runs distinct source labels; do NOT relabel correlated
re-runs as independent (contract §7.1 — over-counting).evidence (optional): a confirmed-real citation.
Converting a /deep-research run → research-claims.json
/deep-research returns result.findings[] (with claim, confidence
high/medium/low, vote, sources) and result.refuted[]. Suggested mapping:
| deep-research | truth_value | confidence |
|---|
| confirmed finding, confidence high | T | 0.85–0.95 |
| confirmed finding, confidence medium | P | 0.6–0.75 |
| confirmed finding, confidence low | U | ~0.4 |
refuted[] entry | F | by vote margin |
Fail-closed (contract §7.2): a multi-judge panel confirms well but catches
invalidity poorly (~96% TPR / <25% TNR). Treat panel agreement as a gate to
promote, and let any credible dissent push to U/D/C — never relabel
correlated judges as independent sources. Verify every cited source resolves
before writing it as evidence (deep-research has hallucinated arXiv ids).
Scheduling
Wrap the command in your scheduled-runner of choice (cron, the
octo:schedule skill, or a state/-driven local runner like the adversarial-qa
runner). Commit state/assumptions/belief-events.jsonl if you want the belief
story versioned; otherwise treat it as regenerable local state.
What it does NOT do
- It does not invoke
/deep-research itself — produce research-claims.json
first (a /deep-research run + the mapping above), then run the loop.
- It does not run tests to verify code assumptions — feed test outcomes as
additional
@ai: annotation reconciliation or research claims.
